Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cross Domain Adaptation using Adversarial networks with Cyclic loss

Manpreet Kaur, Ankur Tomar, Srijan Mishra, Shashwat Verma

TL;DR

This work tackles domain shift in deep learning by introducing a cycle-consistent, adversarial domain adaptation framework for a regression task: steering angle prediction across synthetic source data (Udacity) and real-world target data (comma.ai). The system deploys two cross-domain translators $G_{S \rightarrow T}$ and $G_{T \rightarrow S}$, paired discriminators $D_{S \rightarrow T}$ and $D_{T \rightarrow S}$, and a steering regressor $R_{Steering}$, trained in three phases with a reconstruction (cyclic) loss to preserve ordinal relationships. Phase 1 establishes a solid source-domain regressor, Phase 2 trains the domain translators with GAN objectives and reconstruction losses, and Phase 3 jointly optimizes all networks to achieve cross-domain generalization, yielding a 12.09% improvement in Average Absolute Relative Error on held-out target data. The results indicate that incorporating cycle-consistency and semantic retention in domain translation can improve cross-domain predictions, offering a path toward unsupervised data generation and broader generalization in self-driving perception tasks.

Abstract

Deep Learning methods are highly local and sensitive to the domain of data they are trained with. Even a slight deviation from the domain distribution affects prediction accuracy of deep networks significantly. In this work, we have investigated a set of techniques aimed at increasing accuracy of generator networks which perform translation from one domain to the other in an adversarial setting. In particular, we experimented with activations, the encoder-decoder network architectures, and introduced a Loss called cyclic loss to constrain the Generator network so that it learns effective source-target translation. This machine learning problem is motivated by myriad applications that can be derived from domain adaptation networks like generating labeled data from synthetic inputs in an unsupervised fashion, and using these translation network in conjunction with the original domain network to generalize deep learning networks across domains.

Cross Domain Adaptation using Adversarial networks with Cyclic loss

TL;DR

This work tackles domain shift in deep learning by introducing a cycle-consistent, adversarial domain adaptation framework for a regression task: steering angle prediction across synthetic source data (Udacity) and real-world target data (comma.ai). The system deploys two cross-domain translators and , paired discriminators and , and a steering regressor , trained in three phases with a reconstruction (cyclic) loss to preserve ordinal relationships. Phase 1 establishes a solid source-domain regressor, Phase 2 trains the domain translators with GAN objectives and reconstruction losses, and Phase 3 jointly optimizes all networks to achieve cross-domain generalization, yielding a 12.09% improvement in Average Absolute Relative Error on held-out target data. The results indicate that incorporating cycle-consistency and semantic retention in domain translation can improve cross-domain predictions, offering a path toward unsupervised data generation and broader generalization in self-driving perception tasks.

Abstract

Deep Learning methods are highly local and sensitive to the domain of data they are trained with. Even a slight deviation from the domain distribution affects prediction accuracy of deep networks significantly. In this work, we have investigated a set of techniques aimed at increasing accuracy of generator networks which perform translation from one domain to the other in an adversarial setting. In particular, we experimented with activations, the encoder-decoder network architectures, and introduced a Loss called cyclic loss to constrain the Generator network so that it learns effective source-target translation. This machine learning problem is motivated by myriad applications that can be derived from domain adaptation networks like generating labeled data from synthetic inputs in an unsupervised fashion, and using these translation network in conjunction with the original domain network to generalize deep learning networks across domains.

Paper Structure

This paper contains 21 sections, 6 equations, 15 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Component Networks
  5. Steering Regression Network $R_{Steering}$
  6. Source to Target and Target to Source Domain Translation networks $G_{S \rightarrow T}$ & $G_{T \rightarrow S}$
  7. Discriminator Networks $D_{ST}$ & $D_{TS}$
  8. Loss Functions and Overall Training Methodology
  9. Phase 1 : Training Steering Angle Regression
  10. Phase 2 : Training the Domain Translation and Discriminators in GAN setting
  11. Phase 3 : Training all the networks together
  12. Datasets
  13. Experiments
  14. Phase 1: Training the regressor
  15. Phase 2: Training domain translation network
  16. ...and 6 more sections

Figures (15)

  • Figure 1: Network used for Steering Angle Prediction $R_{Steering}$
  • Figure 2: Schematics for the domain translation network $G_{S \rightarrow T}$
  • Figure 3: Schematics for the discriminator network $D_{S \rightarrow T}$
  • Figure 4: Schematics for the overall phase 3 training. Left side shows the schema with 5 networks with separate discriminators. Right hand side shows a network structure with a unified discriminator; a hypothesis we weren't fully able to test.
  • Figure 5: Some images from the Udacity dataset
  • ...and 10 more figures